Development of Machine-Learning Prediction Programs for Delivering Adaptive Radiation Therapy With Tumor Geometry and Body Shape Changes in Head and Neck Volumetric Modulated Arc Therapy

Abstract

During radiation therapy for head and neck cancer using volumetric modulated arc therapy, excessive dosing or underdosing occurs as a result of the decrease in tumor volume and changes in body weight. Adaptive radiation therapy (ART) is performed when significant changes are observed; however, the decision to implement ART depends on the oncologist's subjective judgment. The purpose of this study was to present objective indicators for ART and develop a program to predict the need for ART. The study included 47 patients in the non-ART group and 21 patients in the ART group with shape changes. Patients who received ART could not be covered with the prescribed radiation therapy dose due to shape changes. For each patient, 1112 6-dimensional lists were created, including the number of irradiations, amount of change in the clinical target volume (CTV), rate of change in CTV, mean oral cavity dose, age, and body mass index. Support vector machine and k-nearest neighbor were used for machine learning. The random number of test data to be extracted varied from 1 to 9, and a mean accuracy score was calculated. These programs could predict the need for ART if the accuracy score was high. The classification accuracy of the list, including the amount of change in the CTV and rate of change in CTV up to 20 fractions, was 0.963 and 0.967 for support vector machine and k-nearest neighbor, respectively. This program predicted the need for ART with more than 90% accuracy based on shape changes over time in cone beam computed tomography analysis for up to 20 fractions. This may provide significant support for objective decisions to implement ART based on the amount of change over time during treatment.